Spartina alterniflora information extraction method and system based on unmanned aerial vehicle remote sensing data
By using the UAV remote sensing data and the UNet deep learning model of the multi-channel attention mechanism, the problem of inaccurate extraction of information in the existing technology is solved, and accurate extraction and boundary recognition of community in complex backgrounds is realized.
Patent Information
- Application Number
- CN202510281094.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-11
- Publication Date
- 2025-06-27
AI Technical Summary
The prior art cannot achieve accurate extraction of the information of mutual flower rice grass, mainly because it presents diverse morphology, texture and phenological characteristics on remote sensing images, with complex and irregular boundaries.
Using a method based on drone remote sensing data, a multi-spectral data set of the UAV community is created by obtaining the UAV remote sensing data in the target area, and a multi-channel attention mechanism UNet deep learning model is used to extract information to determine the spatial distribution information of the UAV community.
The precise extraction and boundary recognition of the mutual flower grass community in small samples and complex backgrounds is achieved, and the extraction accuracy of the mutual flower grass community in different phenological periods is improved.
Smart Images
Figure CN120219957A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of remote sensing monitoring, and particularly to a method and system for extracting Spartina alterniflora information based on unmanned aerial vehicle (UAV) remote sensing data. Background Art
[0002] Spartina alterniflora is a perennial herb of the genus Spartina in the family Poaceae. Due to its strong reproductive ability and adaptability, it has had a serious impact on the coastal ecosystem and is listed as an alien invasive species.
[0003] Existing technologies for remotely sensing and identifying Spartina alterniflora include the remote sensing index method and the object-oriented method. The remote sensing index method uses the spectral feature differences between different vegetations and classifies different plots and vegetations through the operation of characteristic bands. The object-oriented method performs semantic segmentation of different plots through hue, texture, and edge features, so as to achieve accurate identification of the species distribution boundary. However, due to the diverse morphological, textural, and phenological characteristics of Spartina alterniflora in remote sensing images, with complex and irregular boundaries, it poses challenges to the accurate identification and segmentation of Spartina alterniflora. Therefore, the above two methods cannot achieve accurate extraction of its information.
[0004] Therefore, existing technologies need to be further improved. Summary of the Invention
[0005] In view of the above deficiencies in the prior art, the purpose of the present invention is to provide a method and system for extracting Spartina alterniflora information based on UAV remote sensing data, so as to solve the defect that the information extraction methods in the prior art cannot achieve accurate extraction of Spartina alterniflora information.
[0006] In a first aspect, the present application discloses a method for extracting Spartina alterniflora information based on UAV remote sensing data, which includes:
[0007] Obtain UAV remote sensing data of a target area; the target area contains a Spartina alterniflora community;
[0008] Create a multispectral data set of the Spartina alterniflora community based on the UAV remote sensing data;
[0009] Input the multispectral data set into a trained information extraction model to obtain the Spartina alterniflora feature extraction information output by the information extraction model;
[0010] Determine the Spartina alterniflora spatial distribution information according to the Spartina alterniflora feature extraction information.
[0011] Optionally, the step of obtaining UAV remote sensing data of the target area includes:
[0012] Set the flight parameters of the UAV, select a calibration board, and an image repetition rate according to the characteristics of Spartina alterniflora;
[0013] Control the drone to collect data along a preset flight route to obtain drone remote sensing data of the target area.
[0014] Optionally, the multispectral dataset includes: RGB image labels and four single-channel multispectral images;
[0015] The steps of creating a multispectral dataset of Spartina alterniflora communities based on the drone remote sensing data include:
[0016] Perform correction preprocessing on the collected drone remote sensing data to obtain a corrected dataset;
[0017] Create RGB labels for the corrected dataset to obtain RGB image labels and four single-channel multispectral images.
[0018] Optionally, the structure of the information extraction model is a multi-channel attention mechanism UNet deep learning model that integrates the CBAM module.
[0019] Optionally, the training steps of the information extraction model include:
[0020] Establish a training sample set that includes RGB sample labels and multiple single-band data;
[0021] Synchronously input the RGB sample labels and multiple single-band data in the training sample set into a preset multi-channel attention mechanism UNet deep learning model, and train the multi-channel attention mechanism UNet deep learning model to obtain a trained information extraction model.
[0022] Optionally, the information extraction model includes: an encoder, a decoder, and a first CBAM module provided between the skip connections of the encoder and the decoder; both the encoder and the decoder include multiple connected residual blocks, an SE module connected behind each residual block, and a second CBAM module embedded behind each SE module.
[0023] Optionally, the steps of inputting the multispectral dataset into the trained information extraction model to obtain the Spartina alterniflora feature extraction information output by the information extraction model include:
[0024] Combine the RGB image labels and four single-channel multispectral images into an image dataset;
[0025] Input the image dataset into the encoder. Each residual block in the encoder extracts features from the image dataset, and inputs the extracted first feature information into the connected SE module. After the SE module extracts features from the first feature information, it inputs the extracted second feature information into the connected second CBAM module. After processing the received second feature information, the second CBAM module inputs it into the next residual block.
[0026] And the third feature information output from the second CBAM module is synchronously input into the first CBAM module. After processing the third feature information, the first CBAM module inputs the processed fourth feature information into the decoder with skip connection.
[0027] After the feature information input into the residual block and the feature information input into the decoder pass through the feature extraction of the connected SE module, first CBAM module, second CBAM module and residual block in sequence, the Spartina alterniflora feature extraction information is output.
[0028] Optionally, the residual block and the SE module connected behind the residual block include a connected convolutional layer, a channel attention unit, a residual link and a main path addition unit, and an activation output unit.
[0029] Optionally, the Spartina alterniflora body feature extraction information is a binary segmentation image of Spartina alterniflora. The steps of determining the Spartina alterniflora spatial distribution information according to the Spartina alterniflora feature extraction information include:
[0030] Stitch the binary segmentation images of Spartina alterniflora to obtain a feature stitching map of Spartina alterniflora in the target area;
[0031] Perform threshold segmentation on the feature stitching map to obtain the number of pixels containing Spartina alterniflora marks;
[0032] And according to the spatial resolution of the feature stitching map, the number of pixels and the area corresponding to each pixel, statistically obtain the Spartina alterniflora spatial distribution information.
[0033] In a second aspect, the present application provides a Spartina alterniflora information extraction system based on unmanned aerial vehicle remote sensing data, which includes:
[0034] A remote sensing data acquisition module for acquiring unmanned aerial vehicle remote sensing data of a target area; the target area contains a Spartina alterniflora community;
[0035] A dataset creation module for creating a multispectral dataset of the Spartina alterniflora community based on the unmanned aerial vehicle remote sensing data;
[0036] An information segmentation and extraction module, configured to input the multispectral dataset into a trained information extraction model to obtain Spartina alterniflora feature extraction information output by the information extraction model;
[0037] An information statistics module, configured to determine Spartina alterniflora spatial distribution information according to the Spartina alterniflora feature extraction information.
[0038] Beneficial effects:
[0039] The present invention provides a method and system for extracting Spartina alterniflora information based on unmanned aerial vehicle (UAV) remote sensing data. The method includes obtaining UAV remote sensing data of a target area, where the target area contains a Spartina alterniflora community; creating a multispectral dataset of the Spartina alterniflora community based on the UAV remote sensing data; inputting the multispectral dataset into a trained information extraction model to obtain Spartina alterniflora feature extraction information output by the information extraction model; and determining Spartina alterniflora spatial distribution information according to the Spartina alterniflora feature extraction information. The method of the present invention uses a UAV to obtain remote sensing data of the current area, creates a multispectral dataset of the Spartina alterniflora community with multiple time series and multiple scales based on the remote sensing data, and uses an improved multi-channel information extraction model to realize the boundary extraction and area statistics of the area where the Spartina alterniflora community is located in the multispectral dataset, so as to realize the extraction of Spartina alterniflora information under small sample conditions. Description of the Drawings
[0040] Figure 1 is a flowchart of the steps of a method for extracting Spartina alterniflora information based on UAV remote sensing data provided by the present invention;
[0041] Figure 2 is a flowchart of the training steps of the information extraction model in the method of the present invention;
[0042] Figure 3 is a schematic diagram of the principle of information extraction by the information extraction model in the method provided by the present invention;
[0043] Figure 4 is a flowchart of the steps of processing information by the residual block and SE block in the embodiment of the present invention;
[0044] Figure 5 is a comparison diagram of the information extraction effects of the method provided in the embodiment of the present invention and other methods;
[0045] Figure 6 is a schematic block diagram of a system for extracting Spartina alterniflora information based on UAV remote sensing data provided by the present invention. Detailed Embodiments
[0046] To make the objectives, technical solutions and advantages of the present invention more clear and definite, the following further describes the present invention in detail with reference to the accompanying drawings and by way of examples. It should be understood that the specific examples described herein are only for explaining the present invention and are not used to limit the present invention.
[0047] Remote sensing technology has become an important means for large-scale, refined and long-term monitoring of the changes in coastal wetland vegetation communities. The alien invasive species Spartina alterniflora has had a serious impact on the coastal ecosystem due to its strong reproductive ability and adaptability. However, due to the interference of factors such as tidal flats, seagrasses, and mangroves, it is still difficult to remotely sense and identify Spartina alterniflora intelligently, quickly, and over a large area.
[0048] In the prior art, the commonly used remote sensing identification methods for Spartina alterniflora include methods based on remote sensing indices (such as NDVI, EVI, NDWI, GNDVI) and object-oriented methods. The method combining phenological characteristics and spectral differences is currently the main remote sensing index method for extracting Spartina alterniflora. However, in the coastal areas of low latitudes where the temperature is suitable, Spartina alterniflora can spread and grow throughout the year, and there are Spartina alterniflora communities with different phenological periods in the same area. In particular, the newly emerged Spartina alterniflora and the seedlings of artificially planted mangroves have very similar spectral characteristics, and the commonly used pixel-based remote sensing index methods are difficult to identify Spartina alterniflora.
[0049] The object-oriented method performs semantic segmentation of different plots through hue, texture, and edge features, and then realizes the identification of vegetation species. The main methods of semantic segmentation include machine learning and deep learning. In wetland classification, both can be trained through empirical knowledge such as spectral features, texture features, and shape features. Machine learning requires manual extraction of important features to construct an effective classification model, such as support vector machine (SVM), random forest, decision tree, K-nearest neighbor, etc. Deep learning can carry out information extraction tasks for new scenarios based on existing data sets, such as CNN, RNN, and UNet. Due to the strong feature extraction ability of deep learning and its advantages in processing complex data, higher classification accuracy is often obtained in wetland classification with complex categories and irregular boundaries. The deep learning method based on high-spatial-resolution UAV images has a significantly higher overall accuracy than the machine learning method in the classification of coastal wetland plant species. However, the data volume of the deep learning data set is several times or dozens of times that of machine learning. Therefore, establishing a deep learning model based on a small sample size has important application value for the refined classification of wetlands.
[0050] Spartina alterniflora communities are mainly distributed in areas such as bare flats, the edges of mangroves, and tidal creeks, and often grow mixed with mangroves and saline-alkali vegetation. It presents diverse morphological, textural, and phenological characteristics on remote sensing images, with complex and irregular boundaries, posing challenges to accurate identification and segmentation. Therefore, the present invention proposes a U-Net model based on a multi-channel attention mechanism (MCAU-Net), which integrates the CBAM (Convolutional Block Attention Module) module and uses channel attention and spatial attention to enhance the model's ability to capture cross-channel global information and long-range dependencies between pixels, thereby improving the extraction accuracy of Spartina alterniflora communities at different phenological periods under small sample and complex background conditions.
[0051] To overcome the above-mentioned deficiencies, the present embodiment provides a method and system for extracting Spartina alterniflora information based on unmanned aerial vehicle (UAV) remote sensing data. The method includes: obtaining UAV remote sensing data of a target area, where the target area contains Spartina alterniflora communities; creating a multi-spectral data set of Spartina alterniflora communities based on the UAV remote sensing data; inputting the multi-spectral data set into a trained information extraction model, and using the channel attention and spatial attention enhancement modules in the constructed information extraction model to capture cross-channel global information and long-range dependencies between pixels in the multi-spectral data set, extracting Spartina alterniflora feature extraction information, and determining the spatial distribution information of Spartina alterniflora according to the Spartina alterniflora feature extraction information. Thus, the extraction accuracy of Spartina alterniflora communities at different phenological periods under small sample and complex background conditions is achieved.
[0052] The following further elaborates in more detail on a method and system for extracting Spartina alterniflora information based on UAV remote sensing data provided in the present embodiment with reference to the accompanying drawings.
[0053] In a first aspect, the present application discloses a method for extracting Spartina alterniflora information based on UAV remote sensing data, as Figure 1 shown, including:
[0054] Step S1: Obtain UAV remote sensing data of a target area, where the target area contains Spartina alterniflora communities.
[0055] In this step, a UAV is used to collect data from the target area to obtain UAV remote sensing data containing information on Spartina alterniflora communities in the target area.
[0056] Specifically, the drone is equipped with relevant equipment for acquiring remote sensing data, such as an optical camera, an infrared camera or a laser radar. The drone uses the camera or the laser radar to take aerial photos of the target area or hover in the air to collect data. In one embodiment, the optical camera carried by the drone is used to obtain a high-resolution ground image of the target area. The drone usually uses a global positioning system and an attitude sensor to calculate the position and attitude of the camera, and then triggers the camera to take pictures through a trigger. During aerial photography, the camera's viewing angle and aerial photography height can be set according to actual needs.
[0057] In order to use drones to obtain drone remote sensing data of the target area, this step specifically includes:
[0058] Step S11, setting the flight parameters of the UAV, selecting the calibration plate and the image repetition rate according to the characteristics of Spartina alterniflora.
[0059] In order to ensure the accuracy of drone data collection and the safety of flight, Figure 2 As shown, first you need to set the flight route, flight altitude and aircraft direction for the drone in the target area. Then perform the flight based on the set flight route, flight altitude and flight direction.
[0060] The flight route needs to be set so that the images taken by the drone can cover the entire target area. In addition, the flight route needs to avoid no-fly zones or areas with dense buildings to ensure the safety and smoothness of the flight.
[0061] Furthermore, the flight direction of the UAV can be set to an azimuth angle of 45°-135° with respect to the sun direction according to the dual reflection characteristics of the water body of Spartina alterniflora, so as to avoid the pollution of the image by vegetation hot spots and water surface flares.
[0062] When setting the flight parameters of the UAV, you also need to select a suitable calibration blanket or calibration plate (reflectivity of 75%, 50%, 25%) according to the type of target object (target types include: water bodies, vegetation, buildings, etc.), and take a picture of the calibration plate on the ground for remote sensing reflectivity calibration to eliminate the impact of changes in solar radiation.
[0063] The flight altitude determines the spatial resolution of the image. Different flight altitudes can be selected for new seedlings and late-stage communities of invasive species according to the spatial scale and precision requirements of the target object. The spatial resolution calculation formula is: d = (s*H) / (f*10), s is the pixel size (s, in microns), f is the focal length of the lens (in millimeters), and H is the route altitude. In specific implementation, the flight altitude of the drone can be set to 100m, and the spatial resolution can be set to 2.69cm.
[0064] In addition, in order to ensure that the images captured by the drone cover the entire target area, it is also necessary to set up overlapping areas. The overlapping area setting includes setting the heading overlap rate and the lateral overlap rate of two adjacent images. To ensure that there are obvious feature points in two adjacent captured images, in specific implementation, the heading overlap rate can be set to 70%, and the lateral overlap rate can be set to 60%.
[0065] Step S12: Control the drone to collect data along the preset flight route to obtain the drone remote sensing data of the target area.
[0066] When the above step S11 completes the setting of the flight parameters of the drone, the drone can be controlled to perform corresponding remote sensing data collection according to the set flight parameters, so as to obtain the drone remote sensing data of the target area.
[0067] Step S2: Create a multi-spectral data set of Spartina alterniflora communities based on the drone remote sensing data.
[0068] When the above step obtains the drone remote sensing data, since the data collected by the drone is usually raw data, it is necessary to preprocess the drone remote sensing data to obtain more accurate drone remote sensing data. The collected drone remote sensing data is corrected and preprocessed to obtain a corrected data set.
[0069] Specifically, the steps for preprocessing the drone remote sensing data include: geometric correction, radiometric calibration, and atmospheric correction, etc. Geometric correction is: precisely correct the image by selecting ground control points to eliminate geometric distortion. The ground control points can be selected with reference to topographic maps or obtained from the corrected images. The ground control points should have obvious and clear point markers on the image, such as: road intersections, river intersections, etc., and the ground control points should be evenly distributed throughout the image and meet a certain quantity. Radiometric calibration is to calculate the reflectance based on the radiance image of the calibration panel measured on-site and the aerial images taken, and adjust the image through the calculated reflectance to eliminate the non-linear distortion in the image and reduce the image noise, improving the image quality. Atmospheric correction is to eliminate the influence of atmospheric scattering, absorption, etc. on the image and convert the brightness value of the image into surface reflectance. Common atmospheric correction methods include correction based on radiative transfer models. The radiative transfer model can be the 6S atmospheric correction model. According to the 6S atmospheric correction model, the relationship between the surface reflectance ρ and the radiance L can be deduced, such as the formula: Atmospheric correction coefficient X a 、X b and X cIt can be calculated based on the radiance (L1, L2, L3) obtained from aerial photography of three calibration targets and the reflectance (ρ1, ρ2, ρ3) obtained on the ground. The calibration targets are selected as calibration blankets or ground objects with sufficient area and uniform stability (ground objects with sufficient area and uniform stability). The radiance L obtained from aerial photography can be calculated based on the original DN value and the gain and bias coefficients of the UAV lens.
[0070] After correcting the UAV remote sensing data according to the above steps, further processing is performed on the corrected UAV remote sensing data to create a multi-spectral dataset of Spartina alterniflora communities. Specifically, the further operations on the dataset obtained after correction are as follows: Using the Labelme software to draw regions of interest on the image, making RGB labels, and obtaining RGB image labels and four single-channel multi-spectral images.
[0071] Step S3: Input the multi-spectral dataset into the trained information extraction model to obtain the Spartina alterniflora feature extraction information output by the information extraction model.
[0072] After creating a multi-spectral dataset containing RGB image labels and four single-channel multi-spectral images, the multi-spectral dataset is input into the trained information extraction model, and the feature extraction information of Spartina alterniflora is extracted using the information extraction model.
[0073] Specifically, in combination with Figure 2 、 Figure 3 and Figure 4 As shown, the structure of the information extraction model is a multi-channel attention mechanism UNet deep learning model that integrates the CBAM module. This model integrates residual modules, squeeze-and-excitation blocks, and convolutional block attention modules. Through residual learning, it improves the deep feature transfer ability, channel attention enhances feature expression, spatial attention optimizes local and global feature dependencies, and at the same time combines the attention mechanism in skip connections to strengthen the model's ability to model cross-channel global information and long-range dependencies between pixels, thereby improving the U-Net's accurate segmentation and boundary recognition ability for Spartina alterniflora communities at different phenological periods in small-sample, high-complexity backgrounds.
[0074] In one implementation, the training steps of the information extraction model include:
[0075] Step S31: Establish a training sample set containing RGB sample labels and multiple single-band data.
[0076] The training data set includes RGB sample labels and four single-band data (green band, red band, red edge band and near-infrared band). The RGB sample labels in the training data set are selected from images of Spartina alterniflora communities with different phenological periods and invasion degrees to produce sample labels to ensure the universality of the data set. The selected phenological periods include the early and middle growth periods, the late growth period and the senescence period. There are differences in spectral and texture characteristics in different phenological periods. The vegetation index of the community in the growth period is higher (average NDVI>0.5), the community in the early and middle growth periods has less lodging, and the lodging is obvious in the late period. The vegetation index of the community in the senescence period is reduced (NDVI<0.4), and all lodging occurs. The invasion degree is divided into early, middle and late periods. In the early stage of invasion, there are scattered individual communities with a shape close to a circle. In the middle stage of invasion, there are non-circular individual communities. In the late stage of invasion, there are continuous and continuous communities. The four single-band data include: green band, red band, red edge band and near-infrared band. In the specific implementation, software with band extraction function can be selected to extract each single-band data.
[0077] Step S32: synchronously input the RGB sample labels and multiple single-band data in the training sample set into the preset multi-channel attention mechanism UNet deep learning model, train the multi-channel attention mechanism UNet deep learning model, and obtain a trained information extraction model.
[0078] The structure of the information extraction model is an improved U-Net network structure, which incorporates the CBAM module and the SE module. Figure 3 As shown, the information extraction model includes: an encoder, a decoder, and a first CBAM module arranged between the encoder and the decoder in a jump connection; the encoder and the decoder each include a plurality of connected residual blocks, a SE module connected behind each residual block, and a second CBAM module embedded behind each SE module. The feature extraction result output by the encoder is input to the first CBAM module and the decoder in the jump connection, and the final feature extraction result is obtained by the output of the last residual block of the encoder, the SE module, and the first CBAM module.
[0079] When combining the RGB sample labels corresponding to the target area and the four single-channel multispectral images as the image dataset input data and inputting them into the information extraction model, it is necessary to ensure that the size of the input data is consistent in order to perform effective feature extraction and multispectral information fusion.
[0080] Furthermore, it is necessary to match the number of channels and adjust the number of channels of the SE module and the CBAM module to keep it consistent with the number of channels of the input data, so as to optimize the information interaction ability between channels and improve the adaptability of the model to different spectral features.
[0081] The encoder in the information extraction module is improved based on the U-Net encoder. Specifically: After each residual block in the U-Net encoder, an SE module is introduced to enhance channel attention, enabling the information extraction module to adaptively adjust the importance of different channel features. At the same time, a CBAM module is embedded after each convolutional block in the encoder to strengthen the global feature modeling ability while extracting local features, and the residual connection is used to improve the gradient flow to ensure the effective transmission of deep information, thus more accurately capturing long-range spatial dependencies.
[0082] The decoder in the information extraction module is also improved based on the U-Net decoder. By introducing a residual block after the upsampling layer of the U-Net decoder to enhance the robustness of the recovered features, and at the same time, using the CBAM mechanism to process the feature map before the skip connection splicing to enhance semantic information transmission and improve the decoding accuracy, so that the target features can be more completely recovered. In addition, the SE module is incorporated into the decoder to further optimize the feature weights between channels to enhance the attention to the target area.
[0083] Furthermore, in the skip connection part between the encoder and the decoder, a CBAM module is added, and then the feature map output from the CBAM module is spliced with the feature map extracted in the decoder to ensure that the information extraction model can retain and enhance key information during the high-resolution recovery process and improve the accurate recognition ability of the Spartina alterniflora boundary area.
[0084] After the structure of the information extraction model is constructed, the training data set is input into the constructed information extraction model, the constructed information extraction model is trained, and the model parameters are adjusted based on the output results of the model until the model parameters meet the preset conditions.
[0085] Combined Figure 2 As shown, in order to obtain an information extraction model with more accurate extraction results, during the training process, its accuracy is also evaluated. The parameters for accuracy evaluation include: intersection over union, class accuracy, average class pixel accuracy, and average dice coefficient. After the above accuracy evaluation parameters all meet the preset conditions, the information extraction model is trained.
[0086] Furthermore, the steps of inputting the multi-spectral data set into the trained information extraction model to obtain the Spartina alterniflora feature extraction information output by the information extraction model include:
[0087] Step S321: Combine the RGB image label and four single-channel multi-spectral images into an image data set.
[0088] Since the RGB image label is a three-channel color image and the four single-channel spectral images are all single-channel grayscale images, if it is necessary to combine the RGB image label and the four single-channel multispectral images into an image dataset, it is necessary to ensure the consistency of the data. The consistency of the data includes: spatial resolution, geographic coordinates, projection system alignment, and unified image size. In specific implementation, the RGB image label and the four single-channel multispectral images can be combined into a five-channel image array. The data of the three channels of the RGB image are respectively filled into the first three channels of the five-channel image array, and then the data of two bands are selected from the four single-channel multispectral images and filled into the last two channels of the five-channel image array, so as to combine into a five-channel image array.
[0089] Step S322: Input the image dataset into the encoder. Each residual block in the encoder extracts features from the image dataset, and inputs the extracted first feature information into the connected SE module. After the SE module extracts features from the first feature information, it inputs the extracted second feature information into the connected second CBAM module; after the second CBAM module processes the received second feature information, it inputs it into the next residual block; and the third feature information output from the second CBAM module is synchronously input into the first CBAM module; after the first CBAM module processes the third feature information, it inputs the processed fourth feature information into the decoder of the skip connection.
[0090] Step S323: After the feature information input into the residual block and the decoder passes through the feature extraction of the connected SE module, first CBAM module, second CBAM module, and residual block in sequence, the Spartina alterniflora feature extraction information is output.
[0091] When the image dataset is input into the encoder of the information extraction model, it will pass through each module in the encoder for downsampling, attention feature extraction, upsampling, attention feature extraction, etc. in sequence, and finally the Spartina alterniflora feature extraction information output by the decoder is obtained.
[0092] Compared with the semantic segmentation method using RGB true color composite images, the multi-channel based semantic segmentation method used in this embodiment highlights the red-edge reflection characteristics of vegetation by integrating the red-edge and near-infrared bands, and has more advantages in distinguishing the boundaries between mangroves and Spartina alterniflora. This method can also be extended to more bands to improve the application range and accuracy of the model.
[0093] Furthermore, the residual block and the SE module connected behind the residual block include a connected convolutional layer, channel attention unit, residual link and main path addition unit, and activation output unit. Specifically, combined with Figure 4As shown in the figure, in step J1, after the residual block receives the input feature information (which is the extracted feature map in specific implementation), the feature information is input into the first convolutional layer, and the first convolutional layer includes a Conv3×3 + BN module + ReLU activation function. In step J2, after the input feature information is sequentially processed by the 3×3 convolution, BN module, and ReLU activation function, it is input into the second convolutional layer. In step J3, the feature information input into the second convolutional layer is sequentially processed by the Conv3×3 + BN module of the second convolutional layer and then input into the channel attention module. In step J4, after the feature information input into the channel attention module is feature-adjusted, it is input into the residual connection module. In step J5, the feature information after feature adjustment is added to the feature information input in step J1 in the residual connection module and then input into the activation layer. In step J6, the activation layer performs a threshold comparison on the added feature information received and then outputs it.
[0094] The information extraction model outputs Spartina alterniflora feature extraction information, which is the binary segmentation result obtained after feature extraction of the input multi-spectral data set, and its spatial resolution is consistent with the input image. In the test phase, this segmentation result can be further fused with the RGB image to generate a visualization result with Spartina alterniflora markings, ensuring that the name, format, and spatial resolution of the output image are consistent with the original image, making the segmentation result more intuitive and interpretable.
[0095] Step S4: Determine the Spartina alterniflora spatial distribution information according to the Spartina alterniflora feature extraction information.
[0096] Since the Spartina alterniflora feature extraction information output by the information extraction model is the position information of each region containing the Spartina alterniflora community extracted, the spatial distribution information of Spartina alterniflora within the entire target region can be obtained based on the extracted position information of each region.
[0097] Use the multiple regional feature images output by the model to replace the original image, and use a jigsaw software for splicing to complete the production of the entire map; then use ENVI software, through threshold segmentation, obtain the number of marked pixels, and perform area statistics according to the spatial resolution of the image.
[0098] The present invention constructs a high spatial resolution and multi-spectral remote sensing image data set of Spartina alterniflora drones, and proposes an information extraction model of a multi-channel attention mechanism UNet integrating the CBAM module to solve the problems in the task of identifying the spatial distribution of Spartina alterniflora, such as the complex background of wetlands, small sample size, and vegetation phenological changes, accurately extract the boundaries between Spartina alterniflora and other ground objects, and statistically calculate the area of the Spartina alterniflora community, providing technical support for monitoring and controlling the vegetation community invading wetlands and protecting biodiversity.
[0099] To adapt to the information extraction of Spartina alterniflora in different scenarios, a UNet information extraction model integrated with CBAM is established. The model is required to obtain accurate boundaries of Spartina alterniflora communities under small sample conditions (100 samples), without large - area misclassification and omission, and perform area statistics. After verification, the precision evaluation parameters of the model, OA, mIoU, mF1, and DICE, reach 0.92, 0.81, 0.79, and 0.81 respectively, and are superior to the two commonly used semantic segmentation models, FTransUNet and YOLO11, in terms of OA and mIoU. The comparison data is shown in Table 1.
[0100] Table 1
[0101]
[0102] From the comparison results of the precision of different models shown in Table 1 and Figure 5 the information extraction results of different models, it can be obtained that the information extraction model provided in this embodiment can obtain relatively accurate information extraction effects.
[0103] In the second aspect, the present application also discloses a Spartina alterniflora information extraction system based on unmanned aerial vehicle (UAV) remote sensing data. As Figure 6 shown, it includes:
[0104] A remote sensing data acquisition module 610, which is used to acquire UAV remote sensing data of the target area; the target area contains Spartina alterniflora communities; its function is as described in step S1.
[0105] A dataset creation module 620, which is used to create a multispectral dataset of Spartina alterniflora communities based on the UAV remote sensing data; its function is as described in step S2.
[0106] An information segmentation and extraction module 630, which is used to input the multispectral dataset into the trained information extraction model to obtain the Spartina alterniflora feature extraction information output by the information extraction model; its function is as described in step S3.
[0107] An information statistics module 640, which is used to determine the spatial distribution information of Spartina alterniflora according to the Spartina alterniflora feature extraction information, and its function is as described in step S4.
[0108] Furthermore, the remote sensing data acquisition module includes: setting the flight parameters of the UAV, selecting a calibration board and an image repetition rate according to the characteristics of Spartina alterniflora; controlling the UAV to collect data along a preset flight route to obtain the UAV remote sensing data of the target area.
[0109] Furthermore, the multispectral dataset includes: RGB image tags and four single - channel multispectral images;
[0110] The steps of creating a multi-spectral data set of Spartina alterniflora communities based on the UAV remote sensing data include:
[0111] Perform calibration preprocessing on the collected UAV remote sensing data to obtain a calibrated data set; create RGB labels for the calibrated data set to obtain RGB image labels and four single-channel multi-spectral images.
[0112] Furthermore, the structure of the information extraction model is a multi-channel attention mechanism UNet deep learning model that integrates the CBAM module.
[0113] Furthermore, the training steps of the information extraction model include:
[0114] Establish a training sample set that includes RGB sample labels and multiple single-band data; synchronously input the RGB sample labels and multiple single-band data in the training sample set into a preset multi-channel attention mechanism UNet deep learning model, and train the multi-channel attention mechanism UNet deep learning model to obtain a trained information extraction model.
[0115] Furthermore, the information extraction model includes: an encoder, a decoder, and a first CBAM module disposed between the skip connections of the encoder and the decoder; both the encoder and the decoder include a plurality of connected residual blocks, an SE module connected behind each residual block, and a second CBAM module embedded after each SE module.
[0116] Furthermore, the steps of inputting the multi-spectral data set into the trained information extraction model to obtain the Spartina alterniflora feature extraction information output by the information extraction model include:
[0117] Combine the RGB image labels and the four single-channel multi-spectral images into an image data set;
[0118] Input the image data set into the encoder, and each residual block in the encoder extracts features from the image data set and inputs the extracted first feature information into the connected SE module. After the SE module extracts features from the first feature information, it inputs the extracted second feature information into the connected second CBAM module; the second CBAM module processes the received second feature information and then inputs it into the next residual block;
[0119] And the third feature information output from the second CBAM module is synchronously input into the first CBAM module; the first CBAM module processes the third feature information and then inputs the processed fourth feature information into the decoder of the skip connection;
[0120] The feature information input to the residual block and the decoder sequentially passes through the connected SE module, the first CBAM module, the second CBAM module, and the feature information extraction of the residual block, and then outputs the Spartina alterniflora feature extraction information.
[0121] Further, the residual block and the SE module connected behind the residual block include a connected convolutional layer, a channel attention unit, a residual link and a main path addition unit, and an activation output unit.
[0122] Further, the Spartina alterniflora body feature extraction information is a binary segmentation image of Spartina alterniflora; the steps of determining the Spartina alterniflora spatial distribution information according to the Spartina alterniflora feature extraction information include:
[0123] Stitch the binary segmentation images of the Spartina alterniflora to obtain a feature stitching map of the Spartina alterniflora in the target area;
[0124] Perform threshold segmentation on the feature stitching map to obtain the number of pixels containing Spartina alterniflora markers;
[0125] And according to the spatial resolution of the feature stitching map, the number of pixels, and the area corresponding to each pixel, statistically obtain the Spartina alterniflora spatial distribution information.
[0126] The present invention provides a method and system for extracting Spartina alterniflora information based on UAV remote sensing data, by acquiring UAV remote sensing data of a target area; the target area contains a Spartina alterniflora community; creating a multi-spectral data set of the Spartina alterniflora community based on the UAV remote sensing data; inputting the multi-spectral data set into a trained information extraction model to obtain Spartina alterniflora feature extraction information output by the information extraction model; and determining Spartina alterniflora spatial distribution information according to the Spartina alterniflora feature extraction information. The method of the present invention uses a UAV to obtain remote sensing data of the current area, creates a multi-temporal and multi-scale multi-spectral data set of the Spartina alterniflora community based on the remote sensing data, and uses an improved multi-channel information extraction model to realize the boundary extraction and area statistics of the area where the Spartina alterniflora community is located in the multi-spectral data set, so as to realize the information extraction of Spartina alterniflora under small sample conditions.
[0127] Those skilled in the art will readily think of other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses, or adaptations of the present invention, which follow the general principles of the present invention and include known common knowledge or conventional technical means in the technical field not disclosed in this application. The specification and examples are only regarded as exemplary, and the true scope and spirit of the present invention are pointed out by the following claims.
[0128] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or N embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples.
[0129] It can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present application.
Claims
1. A method for extracting information of Spartina alterniflora based on unmanned aerial vehicle remote sensing data, characterized in that: include: Obtain UAV remote sensing data of the target area; The target area contains a community of Spartina alterniflora; Creating a multispectral dataset of Spartina alterniflora communities based on the drone remote sensing data; Inputting the multispectral data set into a trained information extraction model to obtain feature extraction information of Spartina alterniflora output by the information extraction model; The spatial distribution information of Spartina alterniflora is determined according to the feature extraction information of Spartina alterniflora.
2. The method for extracting information of Spartina alterniflora based on unmanned aerial vehicle remote sensing data according to claim 1, characterized in that: The step of obtaining the UAV remote sensing data of the target area includes: According to the characteristics of Spartina alterniflora, the flight parameters of the UAV were set, and the calibration plate and image repetition rate were selected; Control the UAV to collect data along the preset flight route and obtain UAV remote sensing data of the target area.
3. The method for extracting information of Spartina alterniflora based on unmanned aerial vehicle remote sensing data according to claim 1, characterized in that: The multispectral dataset includes: RGB image labels and four single-channel multispectral images; The step of creating a multispectral data set of the Spartina alterniflora community based on the drone remote sensing data comprises: Perform correction and preprocessing on the collected UAV remote sensing data to obtain a corrected data set; RGB labels are produced for the corrected dataset to obtain RGB image labels and four single-channel multispectral images.
4. The method for extracting information of Spartina alterniflora based on unmanned aerial vehicle remote sensing data according to claim 1, characterized in that: The structure of the information extraction model is a multi-channel attention mechanism UNet deep learning model that integrates the CBAM module.
5. The method for extracting information of Spartina alterniflora based on unmanned aerial vehicle remote sensing data according to claim 4, characterized in that: The training steps of the information extraction model include: Establish a training sample set containing RGB sample labels and multiple single-band data; The RGB sample labels and multiple single-band data in the training sample set are synchronously input into the preset multi-channel attention mechanism UNet deep learning model, and the multi-channel attention mechanism UNet deep learning model is trained to obtain a trained information extraction model.
6. The method for extracting information of Spartina alterniflora based on unmanned aerial vehicle remote sensing data according to claim 4 or 5, characterized in that: The information extraction model includes: an encoder, a decoder and a first CBAM module arranged between the encoder and the decoder in a jump connection; the encoder and the decoder each include a plurality of connected residual blocks, an SE module connected behind each residual block and a second CBAM module embedded behind each SE module.
7. The method for extracting information of Spartina alterniflora based on unmanned aerial vehicle remote sensing data according to claim 6, characterized in that: The step of inputting the multispectral data set into a trained information extraction model to obtain the feature extraction information of Spartina alterniflora output by the information extraction model comprises: Combine the RGB image labels and four single-channel multispectral images into an image dataset; The image data set is input to an encoder, each residual block in the encoder performs feature extraction on the image data set, and the extracted first feature information is input to a connected SE module, after the SE module performs feature extraction on the first feature information, the extracted second feature information is input to a connected second CBAM module; after the second CBAM module processes the received second feature information, the second feature information is input to the next residual block; and the third feature information output from the second CBAM module is synchronously input to the first CBAM module; after the first CBAM module processes the third feature information, the fourth feature information obtained by the processing is input to the decoder of the jump connection; The feature information input to the residual block and the decoder is sequentially extracted by the connected SE module, the first CBAM module, the second CBAM module and the feature information of the residual block, and then the Spartina alterniflora feature extraction information is output.
8. The method for extracting information of Spartina alterniflora based on unmanned aerial vehicle remote sensing data according to claim 7, characterized in that: The residual block and the SE module connected to the residual block include connected convolutional layers, channel attention units, residual link and main path addition units, and activation output units.
9. The method for extracting information of Spartina alterniflora based on unmanned aerial vehicle remote sensing data according to claim 7, characterized in that: The feature extraction information of Spartina alterniflora is a binary segmented image of Spartina alterniflora; and the step of determining the spatial distribution information of Spartina alterniflora according to the feature extraction information of Spartina alterniflora comprises: Splicing the binary segmented images of the Spartina alterniflora to obtain a characteristic splicing image of the Spartina alterniflora in the target area; Performing threshold segmentation on the feature mosaic image to obtain the number of pixels containing Spartina alterniflora markers; And according to the spatial resolution of the characteristic mosaic image, the number of pixels and the area corresponding to each pixel, the spatial distribution information of Spartina alterniflora is obtained by statistics.
10. A system for extracting information about Spartina alterniflora based on drone remote sensing data, characterized in that: include: Remote sensing data acquisition module, used to obtain UAV remote sensing data of the target area; The target area contains a community of Spartina alterniflora; A data set creation module, used to create a multispectral data set of Spartina alterniflora community based on the UAV remote sensing data; An information segmentation and extraction module, used for inputting the multispectral data set into a trained information extraction model to obtain feature extraction information of Spartina alterniflora output by the information extraction model; The information statistics module is used to determine the spatial distribution information of Spartina alterniflora according to the characteristic extraction information of Spartina alterniflora.